Skill: MCP-Driven Conversational Analytics

Capability Domain: AI Tooling
Technology Stack: Model Context Protocol, Claude Desktop
Linked Activities: EDA with MCP and Claude Desktop

Recipe

1. MCP server exposes Athena query execution against workgroup ops-lake-analysts, scoped to the contract views in ops_catalog only — never grant it raw-table or cross-database access. There is no fixed set of four GitLab/Jira views.
2. Ask the question in plain language; before trusting the answer, read back the SQL the assistant ran — an unread query is an unverified claim. Use $.payload.* on backfill rows.
3. Use this track first for "what changed" and "does X still hold" questions — then move anything that will inform a real decision into the Pandas/Jupyter track.
4. Write every finding worth keeping into the Claude Desktop MCP Findings Notes artifact in the same session.

Escalate

• Any request to grant this connector write access, or access beyond the contract views — Data Platform Engineer first.

Pitfalls

• Treating a conversational answer as final rather than a hypothesis.